ClaudeにCUDAカーネルを構築させ、オープンモデルを教え込むことに成功！

Back to Articles We got Claude to teach open models how to write CUDA kernels!

Upvote 142

The best thing about agent skills is upskilling your agents on hard problems. There are two ways to look at that:

You can take Opus 4.5 or other SOTA models and tackle the hardest problems out there.

You can take models that run on your laptop and upskill them to harder problems. In this blog post, we’ll show you how to take on the latter.

This blog post walks through the process of using a new tool, upskill

What are agent skills?

In case you missed it, agent skills are taking the coding agent game by storm. In fact, they’re a straightforward concept to define model context as files, like instructions as markdown and code as scripts. The file format makes them easy to generate, share, and review. In short, they’re a practical medium to share capabilities across models and tools, and they're most useful in specific domains or hard problems. Not stuff the model can do well anyway.

This post showcases this process by using Claude to generate a Skill file that can be used by open source models for a complex and specialized task: write CUDA kernels. We first tried a simple skill based on existing documentation, and we found that it improved performance for some others, but not all. In fact, it could even degrade performance or increase token usage for some models. Check out the plot below to see the performance of the model with and without the basic skill.

Now, let's walk through how you can use upskill

1. Get the teacher (Claude Opus 4.5) to build a kernel

First, we use Claude Code to build a kernel interactively and export the trace. We worked through the process by instructing, validating, and adding documentation links. This somewhat naive process is important to reveal the models' initial challenges. In fact, you can iterate on this multiple times, by trying to solve the task with draft versions of the skill, and experimenting with smaller models. Each time, you can instruct the agent to improve the skill and test it on the smaller model.

Here's an example of the skill that we created and have been using to build kernels. We started from this agent trace where the agent was able to build a kernel, but not without some help.

1. Make an agent skill from the trace

Once the teacher model has performed the task, we need them to make a skill. There are a number of effective ways to do this.

Within the same session, instruct the agent to create a skill file for the task it just completed.

Use Anthropic ‘skill creator’ skill either within the agent session or with an exported trace and a new agent session.

Use the upskill

In most cases, the first 2 options result in functional skills. However, the performance of an agent with the skill is unknown. That’s where upskill

1. Take your skill to an open source, smaller, or cheaper model

Finally, we need to transfer our newly created skill to the tool or model we intend to use. Most tools like codex

{agent}/skills/{skill_name}/SKILL.md

In this case, we might want to iterate further on the gpt-oss

upskill generate --from {skill}

There is more to agent skills than model performance. Often agents can reach a given accuracy with or without a skill, they just need to consume more tokens to get there. For recurring tasks, we want to optimize agents to use less tokens to achieve the same accuracy. The results below reveal another dimension to the skill. Some models are significantly reducing their performance token usage, whilst others are using more tokens with the skill. For example, with moonshotai/Kimi-K2-Thinking

tldr; try out and evaluate models with the skills you create. Use upskill eval

That’s the high level end to end of upskilling your coding agents on hard problems. Try out upskill now like this:

install upskill pip install upskill # or use uvx uvx upskill --help # generate a skill based on an agent trace upskill generate "write nvidia kernels" --from ./trace.md # evaluate models on a skill upskill eval ./skills/my-skill/ --model haiku --model sonnet # generate skills for local models upskill generate "parse YAML" --model opus --eval-model "unsloth/GLM-4.7-Flash-GGUF:Q4_0" --eval-base-url http://localhost:8080/v1

Deep dive tutorial into building kernels with agent skills

We have a high level understanding of how we can upskill an agent. Let’s now look at the use case we solved for writing CUDA kernels.

We didn’t just want to write kernel code, but understand the full kernel-builder workflow: project structure, build.toml

The kernel-builder-cuda-kernels

With this skill, you can tell Claude things like:

Build a fused LayerNorm + GELU kernel optimized for H100.

And Claude will create the complete project structure, CUDA implementation, and build configuration—following the exact conventions that kernel-builder expects.

This isn't about generating boilerplate. The skill encodes domain expertise: H100 uses compute capability 9.0, shared memory should be aligned to 128 bytes, async memory copies require __CUDA_ARCH__ >= 900

Setup and Install

Install upskill:

pip install upskill # or use uvx for one-off runs uvx upskill --help

Set your API key:

export ANTHROPIC_API_KEY=sk-ant-... export HF_TOKEN=hf_...

That's it. Upskill uses Anthropic Claude Opus-4.5 model by default but also supports OpenAI and local models via OpenAI-compatible endpoints as generators. We want to use the more expensive and higher quality models to generate skills, and the smaller ones to use them. Think robin hood.

Skill Generation

Let's walk through generating a skill that teaches agents how to build CUDA kernels with HuggingFace's kernels

Generate the Skill

Start with a clear task description:

upskill generate "build optimized CUDA kernels for PyTorch using HuggingFace kernel-builder"

Above we used upskill, but it could in fact be any agent or chat tool and an exported trace.

upskill generate "write kernels" --from <agent-trace>.md

Also, we could start from an existing skill and add to it:

upskill generate "add more error handling and edge cases" --from ./skills/kernel-builder-cuda-kernels/

upskill loads the existing skill, applies your improvements, and re-evaluates to ensure the changes help.

upskill creates a skill, generates test cases, evaluates performance, and refines based on failures:

Generating skill with sonnet... Generating test cases... Evaluating on sonnet... (attempt 1) 60% -> 95% (+35%) OK kernel-builder-cuda-kernels Build optimized CUDA kernels for PyTorch using HuggingFace kernel-builder. SKILL.md ~520 tokens baseline ████████████ 60% with skill ███████████████████ 95% (+35%) Saved to ./skills/kernel-builder-cuda-kernels

The baseline shows how the model performs without any skill. The "with skill" result shows performance after the skill is injected into context. A 35% improvement means the skill is working.

The skill is saved as a directory following the Agent Skills specification:

./skills/kernel-builder-cuda-kernels/ ├── SKILL.md # Main instructions (~520 tokens) └── skill_meta.json # Metadata and test cases

--- name: kernel-builder-cuda-kernels description: Build optimized CUDA kernels for PyTorch using HuggingFace kernel-builder. --- # Building CUDA Kernels with kernel-builder ## Overview This guide explains how to create optimized CUDA kernels for PyTorch models using HuggingFace's kernel-builder. It covers project setup, kernel implementation, and building for specific GPU architectures like NVIDIA H100. ## Project Structure project/ ├── build.toml # Build configuration ├── kernel_src/ # CUDA kernel implementations │ ├── attention.cu │ ├── layernorm.cu │ └── geglu.cu └── torch-ext/ # PyTorch C++ bindings └── torch_binding.cpp ## Build Configuration Create build.toml to define your kernel package: [general] name = "diffuser_kernels" backends = ["cuda"] [general.cuda] # H100 is compute capability 9.0 capabilities = ["9.0"] ...

Evaluate on a Different Model

The important test is: does this skill help local or cheaper models to build kernels?

Start a local OpenAI-compatible server with a web UI: llama-server -hf unsloth/GLM-4.7-Flash-GGUF:Q4_K_M # Evaluate on local model (llama.cpp server) upskill eval ./skills/my-skill/ --model "unsloth/GLM-4.7-Flash-GGUF:Q4_0" --base-url http://localhost:8080/v1

Generating skill with sonnet... Generating test cases... Evaluating on "unsloth/GLM-4.7-Flash-GGUF:Q4_0"... (attempt 1) 40% -> 85% (+45%) OK baseline ████████░░░░░░░░░░░░ 40% with skill █████████████████░░░ 85% (+45%) Saved to ./skills/kernel-builder-cuda-kernels

A 45% improvement on "unsloth/GLM-4.7-Flash-GGUF:Q4_0"

This is the core value proposition: use expensive models to create skills, then deploy those skills with cheap or local models.

How the evaluation in upskill works

upskill uses a teacher-student approach to evaluate models where the teacher model generates test cases for the student model to be evaluated on.

Teacher model (Opus) generates the skill

Test cases (Opus) are generated automatically from the task description

Student model (local) is evaluated with and without the skill

Skill lift measures the improvement

If you pass an existing skill to upskill eval

{ "cases": [ { "input": "Create a build.toml for a CUDA kernel targeting H100", "expected": {"contains": "9.0"} }, { "input": "Write a basic CUDA kernel template with proper includes", "expected": {"contains": "cuda_runtime.h"} } ] }

We can also test how a skill performs across different models:

upskill eval ./skills/kernel-builder-cuda-kernels/ --model haiku --m kimi --runs 5

Evaluating kernel-builder-cuda-kernels across 2 model(s) 3 test case(s), 5 run(s) per model haiku Pass rate: 4/5 (80%) Avg assertions: 2.8/3 sonnet Pass rate: 5/5 (100%) Avg assertions: 3.0/3 ┏━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓ ┃ Model ┃ Pass Rate ┃ Avg Assertions ┃ Avg Tokens ┃ ┡━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩ │ haiku │ 4/5 │ 2.8/3 │ 1250 │ │ kimi │ 5/5 │ 3.0/3 │ 1890 │ └────────┴───────────┴────────────────┴────────────┘

This helps you find the cost-performance sweet spot: maybe Haiku with the skill is good enough for your use case, saving significant API costs.

We've shown that upskill can create validated skills that transfer domain expertise from powerful models to cheaper ones. The kernel-builder skill is just one example of what's possible.

Some things to try:

Generate skills for your internal tools

Build a skill library for your codebase

Capture tribal knowledge

Benchmark across models

The approach works for any specialized task where you'd otherwise write detailed prompts repeatedly. Skills are portable across Claude Code, Codex, Cursor, and other tools that support the Agent Skills specification.

Agent Skills Specification

HuggingFace kernel-builder